I have not used ORCA myself yet. However, maybe a more general explanation helps too. Assuming you mean a Geometry Optimization using DFT, some contributions to the electronic energy (exchange-correlation term) and the resulting gradients are computed via numerical integration on a grid. The denser/larger the grid the more accurate is the result compared to the 'exact' analytical evaluation of these contributions.

Thus, DFT programs usually require selection of a grid size, e.g., Grid3. How exactly grid3 is defined for ORCA, should be found either in the manual or in the corresponding literature. I normally use the program package Turbomole, in which grid size 3 is the default option (but attention, I do not know, if Grid3 of ORCA is the same).

A too small grid may result in convergence but wrong answers or even yields convergence problems in the optimization. There are techniques for small grids (keyword: 'weight derivatives'), which correct the 'grid effect' in the gradients and yield good convergence even for small grids and even though the energy hypersurface is poor in terms of physical accuracy. Most DFT programs have weight derivatives implemented but not necessarily switched on by default.

If you have electronic/structure convergence problems in a broken symmetry calculation (e.g. as approximation for an open shell singlet state), a poor grid size is usually not the dominant problem. However, it depends on the electronic situation of your system, if a default grid is enough. The best way to find that out is to run Grid3 and even a larger grid size on the same test system and check wether the results differ significantly or not.

I'd like to add some comments to the answer given before. The definition of the grids are given in the ORCA manual. As there is no general answer which grid to use because of too many factors that are affected by the grid, I would recommend to test it with several settings. By the way if you also use the RIJCOSX approximation you should also take care of the "Gridx" keyword.

Without any further information it is hard to say if your convergence problems are caused by the grid settings. I would like to invite you to the ORCA forum (http://www.cec.mpg.de/forum/) where all the experts from the ORCA team will help you. We would also be able to look into your input and output files and help you to solve your problems.

Dear Dr Susanne, Thanks indeed for all. I am sending you one UKS .inp file where unrestricted HF/UKS for GO was done for only one single spin. The calculation is already more than 48 machine hrs in a HPZ200 workstation with a support of 8 GB RAM and 1 TB hard disk support service. The calculation is already completed more than 15 Geometry optimization cycle but no way it is converging and one more thing during SCF iterations it is restarting incremental FOCK matrix and resetting DIIS (as per input file)...To me it seems there is somehow under a false loop where from it cannot give the final single point energy, though I can see the single point energy after each optimization. No way converging..Looking forward for a proper solution, what exactly creating problem...

Thanks for all information..

Thanks for the input file. What I saw directly is that you used the RIJONX keyword which I don't have experience with. What you can do is using the RIJCOSX aprroximation instead of RIJONX which is described in the manual. Another thing that you can look into is the structure at every step of the optimisation. Do they look useful? Is the energy lowered in each step? Did the SCF converge for each step? How did the SCF converge? If you don't mind you can also send me the output file. I would also recommend to do a test calculation with a larger grid (maybe Grid6 or something like that) What woudl happen if you would skip the %scf block? Did you look into the orbitals?

Just one more point...

Energy calculations are less sensitive to the grid precision. But if you perform an optimization, you need gradients. And you need gradients that are really the gradients of the energy, otherwise optimization algorithms fail or are less efficient. To get accurate gradients you need a fine grid. I would even recommend grid4.

If I am not wrong, gradients are more accurate with RIJCOSX than with RIJONX. Check the manual.

Hi Dr. Crehuet......I am trying to redo following your comments...Thanks

Hello Dr. Susanne..Unfortunately no out file was generated yet. I am trying with Grid 4 and RIJCOSX. Lets see. Thanks indeed.

Hi once again.. I have just tried the calculation with removal of flag %scf block...the run is interrupted..The modified inp file is now as follows..

The input file you uploaded still has an %scf block. So maybe I misunderstood your post. Would you please upload the output file? It would it make easier to see why the calculation is interrupted.

By the way, if you use RIJCOSX you should also increase the Gridx to Gridx5 or Gridx6 for example.

Here is the modified input file....Please have a look. Thanks

What is your error message? Is it the SCF convergence?

Dear Dr Banerjee,

I've run your input file in my workstation with "Grid4 NofinalGrid" without any problem neither of optimization or SCF convergence (I also removed the "normalprint" keyword which also adds a lot of output). I attach the gzipped output file (where you can also check the input lines). Everything looks normal in the optimization.

Hope this helps.

Dear Dr. Crehuet......It's a great help....In my computer it is still going on with Grid 6 NofinalGrid and with Normal print flags...

Dear Sir/Ma,

do you have experience about the Gamess to do DFT geometry optimization? I'm doing it, but I have the scf convergence problem. Does anyone can help me? Thank you very much!

Best wishes,

Feng